The first reaction to a $200/mo bundle, in any market, is people rebuilding it themselves. Grok Bot launched in beta on August 11, 2026, and within days the open-source community had working clones of the core pattern — persistent agents, each with its own computer, running against your own model API key instead of a subscription. This is what actually exists, what it costs in time rather than money, and where it falls short of a hosted product.
Every repository, license, and commit date below was checked twice: once against the live GitHub page, once by cloning the repository directly over git to confirm the commit hash independently. Where a project turned out abandoned, a plain fork, or unverifiable, it's named and ruled out rather than quietly dropped.
What you need to assemble
"Self-host a Grok Bot alternative" isn't a single install — it's four separate pieces that need to talk to each other.
A VM or sandbox layer. Somewhere isolated for the agent to actually run — a container or virtual machine with its own filesystem, so one agent's mess doesn't touch another's and a crashed run doesn't take down your host machine.
A computer-use or browser layer. The part that actually drives a browser or desktop — clicking, typing, reading the screen — the same capability Grok Bot and CloudAxis both build on top of.
An orchestration layer. Scheduling, triggers, and the workflow logic that decides when an agent runs and what it does next — the cron-plus-logic layer, not just a one-shot script.
Model access via your own API keys. Every piece above needs a reasoning model behind it, and in a self-hosted stack that's billed directly to you by whichever provider you pick, metered per token, with no bundle absorbing the cost.
A hosted product bundles all four behind one login. Self-hosting means picking one project per layer and wiring them together yourself.
The projects
Community Grok Bot clones
OpenMausBot — github.com/milind-soni/OpenMausBot — Apache-2.0 — active, 563 commits at time of writing. Built explicitly as an open-source alternative to Grok Bot: a chat-app interface where each bot has its own model, memory, and personality, running on a cloud desktop, a local VM, or your own machine. It uses the trycua/cua computer-use layer underneath (see below) and Composio for third-party app integrations. This is the closest thing to a direct clone that currently exists, and it's genuinely early — expect rough edges consistent with a project a few days old.
Rakazo — github.com/elie222/rakazo — Apache-2.0 — active, 129 commits at time of writing. An open-source platform for "persistent AI teammates" with web, desktop, and mobile clients. Supports multiple sandbox backends (Docker, E2B, Daytona, local) and includes voice mode and task delegation between bots. Also early, also community-run — check its issue tracker for open bugs before committing real workflows to it.
One project we ruled out: a fork of OpenMausBot under a different account had zero independent commits and no changes beyond the original — not a distinct project, just a copy, so it's excluded here.
Computer-use and browser-automation frameworks
This is the most mature layer of the stack by a wide margin — these projects predate Grok Bot and are what both OpenMausBot and most hosted agent products (CloudAxis included) are ultimately built on top of.
browser-use — github.com/browser-use/browser-use — MIT — commit from 2026-08-19. The most widely adopted open-source browser-driving layer for AI agents; connects a reasoning model to a real browser session.
Stagehand — github.com/browserbase/stagehand — MIT — commit from 2026-08-20. Browserbase's framework for combining scripted browser automation with AI-driven steps where a fixed script would break.
Skyvern — github.com/Skyvern-AI/skyvern — AGPL-3.0 — commit from 2026-08-20. Vision-based browser automation built for workflows like form-filling that resist brittle selector-based scripts. Flag the license before bundling this into anything you plan to offer as a hosted service — AGPL has copyleft obligations that MIT and Apache-2.0 don't.
Anthropic's computer-use reference — github.com/anthropics/anthropic-quickstarts — MIT — commit from 2026-08-19. The public reference implementation for driving a full desktop (not just a browser) from a model, including self-hosted sandbox Docker setups.
cua — github.com/trycua/cua — MIT — commit from 2026-08-20. A computer-use driver layer for cloud desktops, local VMs, and containers — the specific dependency OpenMausBot builds on, useful to know if you're assembling a stack piece by piece rather than taking a clone as-is.
Orchestration and scheduling
n8n — github.com/n8n-io/n8n — commit from 2026-08-21. Licensing needs a closer read than "open source" implies: it's the Sustainable Use License, not an OSI-approved license — content on branches other than the main line is unlicensed, and any file with .ee. in its name requires a paid Enterprise License. The rest is free to self-host and modify.
Dify — github.com/langgenius/dify — commit from 2026-08-21. Also not plain Apache-2.0: it's a modified version that permits commercial use but blocks running Dify itself as a multi-tenant hosted service without a separate commercial license, and requires keeping its branding intact in the frontend.
Both are unambiguously active projects — commits within hours of this research — and both work as the scheduling and workflow layer that turns a one-shot agent script into something that runs unattended on a timer.
VM and sandbox providers
E2B — github.com/e2b-dev/e2b — Apache-2.0 — commit from 2026-08-20. Actively maintained, no deprecation notices, straightforward Apache-2.0 terms — the cleanest option in this category right now.
Daytona — github.com/daytonaio/daytona — Apache-2.0, frozen at tag v0.190.0. This one needs a direct correction to how it's usually described: as of June 2026, Daytona's open-source repository was reduced to a README notice — core development moved to a private codebase, and the public repo "will receive no further updates, fixes, or releases." It remains usable and forkable under its license as-is, but it is not a maintained option going forward. We confirmed this by cloning the repository directly rather than trusting the GitHub page summary — the last real commit is from 2026-06-25, and the working tree now contains only the README and an assets folder.
The real cost of self-hosting
None of the above is free in any meaningful sense — it trades a subscription for a different set of costs, and for a lot of readers those costs are the right trade. Being honest about what they are:
Infrastructure you now run. A VM or sandbox provider bills you directly, usage-based, with no one else absorbing the spikes.
Your own API keys, metered. Every model call is billed to your account directly — no bundle smoothing it out.
Upkeep when a site changes. Browser automation breaks when a target site redesigns a form or adds a bot check. In a hosted product, someone else's team patches it. Here, it's a GitHub issue you file — or fix yourself.
No support line. Community projects don't have an SLA. Fixes land on the maintainers' schedule, and for the two Grok Bot clones specifically — both genuinely early — that schedule is unproven.
The security surface is now yours. Credential storage, session isolation, and audit logging all become things you design, not things a vendor already solved. If you're self-hosting an agent that logs into real accounts, that's the actual job, not a footnote.
When hosted still wins
Self-hosting is the right call for plenty of readers of this post — technical, comfortable owning infrastructure, and specifically wanting the control that comes with it. Hosted still wins for a few clear cases: non-technical users who don't want to run a Docker stack, teams that need billing they can predict in advance rather than a metered API tab, and anyone who doesn't want to personally own the credential-isolation and audit-logging problem described above.
In the interest of full disclosure, CloudAxis is a hosted option in that category — scheduled specialists with a persistent, per-user isolated cloud browser and hard monthly caps from $19/mo, covered in more depth in our pillar guide to the category. Our honest limitation here: we don't offer the visible-repo, inspect-every-line control that self-hosting gives you, and if that's specifically what you want, one of the projects above — not a hosted product, ours included — is the right answer.
Frequently asked questions
Is there an open-source Grok Bot?
Not an official one from xAI, but community projects launched within days of Grok Bot's August 11, 2026 beta that replicate the pattern. OpenMausBot and Rakazo are the two verified, actively maintained examples as of this writing — both early and community-run, not production-hardened products.
Can I self-host AI agents with computer use?
Yes. You need a VM or sandbox layer, a computer-use or browser layer, an orchestration layer for scheduling, and your own model API keys. None of this ships as a single install — you're assembling a stack.
What does it cost to self-host?
No subscription, but real costs elsewhere: your own compute, your own metered API spend, and your own time for setup and ongoing maintenance when something breaks. There's no vendor support line.
Do I need my own API keys?
Yes, for every project in this stack — self-hosting shifts model access from a bundled subscription to metered billing under your own account with whichever provider you choose.
Related reading in this cluster
- Pillar guide: AI agents with their own computer — the full category, hosted and self-hosted alike
- CloudAxis vs Grok Bot
- Grok Bot alternative: always-on AI agents without the $300/mo bundle
- Grok Bot and shared credentials: what multi-agent security actually means
- Grok Bot vs ChatGPT Agent: which actually does your work?
- Is Grok Bot worth it? — what early users report
Want the pattern without assembling the stack yourself?
Put my work on autopilot →